Social Diversity and Duverger

Evidence from South African Local Elections

Karen Ferree University of California, San Diego keferree@ucsd.edu

Clark Gibson University of California, San Diego ccgibson@ucsd.edu

Barak Hoffman Georgetown University bdhoffman@ucsd.edu

Paper prepared for the December 2007 meetings of the Working Group in African Political Economy at Stanford University

Abstract

The effect of institutions and social diversity on political competition is a lively source of debate. Many existing studies on this subject have trouble separating how each affects the number of parties since social cleavages can affect the types of institutions that countries adopt. In this paper, we advance research in this area by examining how political institutions and diversity affects the number of parties at the local level in South Africa. South Africa is an excellent case for examining the independent effect of these factors on the party competition. First, since municipal boundaries and local electoral rules are exogenous to local exigencies, we do not encounter problems of how cleavages affect institutions. Second, South Africa employs a mixed member electoral system at the local level, with voters simultaneously casting a single member vote and a PR one for their local election. Our results suggest that racial diversity influences the number of parties regardless of institutional rules. Institutions, by contrast appear to have no strong effect on the number of parties.

Initiating a long and venerable line of research in comparative politics, Duverger (1952, 1963) argued that electoral institutions exert a powerful effect on the number of political parties: majoritarian systems tend to produce just two parties, while proportional systems allow a wider range of parties to flourish. A succession of authors have refined and built upon these basic intuitions, extending and formalizing his results for a broader array of institutions and further developing the interactive effects between institutions and social diversity implied in his work. It is now commonly understood that electoral institutions mediate the effects of social diversity on the number of parties in party systems: diversity has little to no effect in majoritarian systems, but potentially larger effects in proportional ones. As important as these ideas are, most empirical tests of Duvergers intuitions have been flawed. Many suffer from problems of endogeneity: political traditions shape institutional choice even as institutions shape political traditions, yet the research design employed in most studies does not allow authors to tease out reciprocal flows of causality. Furthermore, institutional theories about the number of parties are most relevant at the district not national level, yet most tests employ national level data, which may obscure empirical patterns evident at a more disaggregated level. Many previous tests have also been hampered by a relatively small number of data points, typically around 100 but sometimes far fewer. And finally, many existing tests combine apples and oranges: they operationalize social diversity as ethnic diversity and then pool together countries where ethnicity is highly salient with countries where it is not salient at all. A more robust test of the importance of social cleavages and institutions on a party system would call for a case with a mixed electoral system at the district level within a country, allowing a direct comparison of different institutions within the same social setting.

This is precisely the case that we exploit in this paper.

We circumvent conventional

problems by utilizing a unique dataset of local elections in South Africa. South African local elections use a mixed electoral system in which voters cast both a proportional representation (PR) vote for an at-large local district as well as a ward vote (akin to a single member plurality, or SMP) vote. By matching these sets of voting results (PR and SMP) with census data on ethnic and racial diversity for nearly 4000 wards, we run a series of very powerful tests on the relationship between the effective number of parties (ENP), institutional rules, and social diversity that avoid many of the problems afflicting earlier studies: all of South Africas municipalities have the same mixed-member institutional design, hence the choice of local institutions does not reflect local political dynamics, eliminating the problem of endogeneity. Our unit of analysis is the ward, akin to the district in Duvergers work and the proper level of analysis. Moreover, we have far more data to work with than cross-national studies, and the nature of social cleavages and their salience are comparable across our units of analysis. Our tests produce some surprising results. We find that institutions exert little to no effect on the number of parties in South African local elections. Ward ENP is very similar regardless of institutional rules. Furthermore, elections under SMP rules produce many violations of Duvergers Rule (party systems where ENP exceeds two). And perhaps most interesting, we find robust evidence that racial diversity produces powerful and consistent results across institutional rules: regardless of whether rules are SMP or PR, more diversity produces more parties. We also find, contrary to recent work by Dickson and Scheve (2007), no evidence of non-linearities in this relationship. Finally, we explore the effects of South Africas nested group structure on the party system. We find that while racial diversity produces more parties, ethnic diversity (controlling for racial diversity) actually decreases the number of parties.

Although our ethnic diversity results are less robust than our racial diversity results, they suggest the value of looking at different types of fractionalization and their effects on party systems. Altogether, our results indicate that social diversity induces violations of the conditions necessary for majoritarian institutions to restrict party number. We discuss two such violations: identity voting and lack of common beliefs about the rank ordering of parties. Social diversity, when it is associated with one or more of these conditions (as we believe is the case in South Africa), interferes with strategic voting, producing failures of coordination, and violations of Duvergers Law. Most prior work, by focusing only on how social diversity increases demand for parties while ignoring its effects strategic behavior, fails to anticipate these effects. 1. Institutions, Ethnic Diversity, and Party Systems: Some Hypotheses from the Literature Beginning with Duverger (1954), political scientists have recognized that party systems reflect both institutional and social factors. In the analysis below, we pull together several hypotheses from this literature. We discuss institutional theories first (focusing on district size), and then move on to explanations focusing on the interaction between social diversity and institutions. a. Institutions and Party Systems A well developed strand of literature in comparative politics addresses the effects of electoral rules (specifically district size) on party systems. Without engaging in an extensive review of this literature (see Riker 1976 and 1982, Lijphart 1994, Cox 1997, and Clark and Golder 2006 for this), we note its primary intuitions. Duverger (1954) argued that due to both mechanical and strategic effects, single-member plurality (SMP) electoral rules discourage the formation of more than two parties. By mechanical effects, he referred to the negative effects

of disproportionality (or a high threshold for converting votes to seats) on small parties. Under SMP rules, it is possible for a party to win a substantial number of votes overall, but failing to win a plurality in any district, not gain representation in the legislature. The strategic effect involves the behavior of candidates and voters. Understanding the mechanical effect of SMP rules, candidates decide to compete in SMP elections only when they believe they have a good chance of winning. Furthermore, voters behave strategically, avoiding candidates they do not believe can win, even if they prefer those candidates more. Altogether, the mechanical and strategic effects of SMP rules depress the number of parties in single member plurality systems, generally capping the equilibrium number at two. Duverger also argued that proportional representation (PR) systems attenuate the mechanical effect and reduce incentives for strategic behavior by candidates and parties, which implies that the electoral system places fewer constraints on the number of parties. This does not mean, as Clark and Golder (2006) point out, that these systems will necessarily have a large number of parties, only that more parties are possible. Cox (1997) builds on Duverger, spelling out the specific conditions under which Duvergers logic should hold and generalizing it to all varieties of electoral rules. According to Cox, voter coordination requires four very specific conditions (see pp. 76-80). First, voters cannot be indifferent between candidates: they must have a strict preference ordering. If they are indifferent between the first and second candidate, they have no incentive to abandon their (losing) preferred candidate. Second, situations where one party will win the election with certainty also reduce incentives to behave strategically: if coordinating has no effect on the outcome, why abandon a losing first choice? Third, voters must be short-term instrumentally rational. If voters are expressive derive benefits merely from the act of voting, regardless of

the outcome or if they care about influencing long term outcomes and are willing to lose in the short term to do so they are not likely to engage in strategic voting. Finally, the ranking of the candidates must be common knowledge: all voters must have a common understanding of the order in which candidates are likely to finish. If these conditions are in place, then voters will concentrate their votes on the candidates who have a chance at winning. Duverger showed that this is two candidates in SMP elections. Cox extends this, showing that the maximum number of candidates (parties) in any system (assuming successful coordination) is the size of the district plus one, the now ubiquitous M + 1 rule. It is important to note, similar to Duverger before him, Cox does not imply that the number of parties will be M + 1, only that the upper limit on party number will be capped at this point. Coxs assumptions are more restrictive than sometimes appreciated. For example, he suggests that they may be less likely to hold in new democracies. The fourth condition common beliefs about the ordering of candidates may be especially problematic where prior electoral patterns provide no guide, parties are in formation, and public polls are scarce or unreliable. We might therefore expect to see more failures of coordination in early elections versus later ones. Moser (1999) finds evidence to support this notion for Russian elections and Clark and Golder (2006) show that the effects of institutions on party systems are not as pronounced or as predictable in new democracies as they are in old ones. Despite these complications, tests of Duvergers theories have strong empirical support. According to Taagapera and Shugart (1993: 455), . . . if one had to give a single major factor [for the number of parties], it would be district magnitude. Amorim Neto and Cox (1997), Benoit (2001), Benoit (2006), Taagepera (1999), and Taagapera and Shugart (1993) have extensive reviews of these studies.

From this literature, we extract four related and uncontroversial hypotheses about the relationship between institutions and the number of parties in the party system.1 Hypothesis 1: Holding constant other factors, the effective number of parties should be higher under PR systems than under SMP systems. Hypothesis 2: The effective number of parties should correlate with the upper bound of the district (M+1, where M is the number of legislative seats to be filled). Hypothesis 3: Under SMP, the effective number of parties should be capped at 2. Under PR, the effective number of parties should not be higher than the district bound (M+1, where M is the number of legislative seats to be filled). Hypothesis 4: Hypotheses 1-3 are more likely to hold in later elections. The first three hypotheses are conditional on the assumptions laid out by Cox. Hypothesis 4 acknowledges the possibility that early elections may fail to satisfy these assumptions and thereby may not conform to Duvergers logic. b. Institutions and Social Cleavages As a recent article by Clark and Golder (2006) points out, Duverger was not simply an institutionalist. He speculated that social forces create the demand for political parties, which political institutions then mediate: where demand exists for multiple parties, and where institutions permit it, we are more likely to see large party systems. In contrast, where demand can be satisfied by a small number of parties, or where institutions are constraining, we should

We follow the predominate trend in the literature and use effective number of parties (ENP) as our operationalization of party number. ENP, due to Laakso and Taagepera (1979), is an inverse Hirschman-Herfindahl index that measures not only the number of parties in the system, but their concentration.

expect a small number of parties. Several more recent studies (Powell, 1982; Ordeshook and Shvetsova, 1994; Amorim, Neto and Cox, 1997; Mozaffar, Scarritt, and Galaich, 2003; Clark and Golder, 2006; Brambor, Clark, and Golder, 2006) have further developed this idea, so much so that scholars commonly understand party systems to reflect the interactive effects of social diversity and institutions.2 Multiple empirical studies support this understanding: using cross national datasets, Ordeshook and Shvetsova (1994) and Amorim Neto and Cox (1997) find that social diversity only affects ENP in permissive systems. Clark and Golder (2006), using an updated dataset and more carefully specified models, produce the same result (which holds especially well for older democracies). Mozaffar, Scarritt, and Galaich (2003) provide a dissenting voice, suggesting that different dynamics hold in Africa: more diversity, perversely, produces fewer parties, as does higher district magnitude. However, Brambor, Clark, and Golder (2006) correct errors in Mozaffar, Scarritt, and Galaich s specification, coding, and interpretation to produce more conventional results from the same data. Scholars have also begun to look at these relationships by using subnational data (Geys 2006, Vatter 2003, Lago Penas 2004) and have by and large confirmed the intuitions of the national studies. In sum, there appears to be a theoretical and empirical consensus that social diversity affects party systems only when institutions allow it.
2

While the institutional side of this equation is relatively well understood, with clear models of voting behavior and candidate entry, the social side is less well developed. The general idea seems to be that voters in social groups have a common desire for a party that exclusively represents their own group, hence demand for parties increases in the number of groups. The behavioral origins of this demand remain obscure in Duverger and his followers but we can draw intuitions from related literatures. Perhaps they lie in expressive instincts (Horowitz 1985). Perhaps voters use the group credentials of the party as a shortcut for making decisions (Mattes, 1995; Chandra, 2004; Ferree, 2006) or expect greater patronage benefits from parties led by members of their own group (Chandra, 2004; Posner, 2005). Regardless of the individual motivations, Duvergers logic requires that voters living under SMP rules caste aside their desire for same-group parties and merge votes with members of other groups if it appears that voting for samegroup parties would require wasting a vote. Demand for same group parties, in other words, is secondary to strategy. This implies a certain flexibility to behavior, a deference to strategic considerations, that may be more consistent with some models of group voting over others. However, these implications are not fleshed out in current work.

10

Dickson and Scheve (2007) construct a different theory to explain how social diversity and institutions interact to shape party systems. They model voters and candidates as having both instrumental and expressive goals. Individuals vote for parties closest to their policy preferences, conditional on not taking action that harms the electoral performance of their group. Candidates can also win awards for gaining office or for being the most electorally popular candidate from her own group (though losing the election). Building on Osborne and Slivinskis (1996) citizen-candidate model, they derive a number of propositions showing that social diversity can affect the equilibrium number of parties even in restrictive (SMP) systems. More specifically, they predict that above a certain demographic threshold (defined by the size of the largest group, where areas with very dominant groups are above the threshold and areas with relatively balanced groups are below it)3, restrictive systems can support more than two parties. The intuition is that in cases with a dominant group (i.e. comprising a large majority of the voters in the electoral district), same group competitors may enter the race without fear that doing so will cause the group to lose. In contrast, evenly matched groups act as a deterrent to same group competitors. Therefore, areas with dominant groups (above the Scheve- Dickson threshold) may have more than two parties, while areas with evenly matched groups (below the Scheve-Dickson threshold) are more likely to conform to Duvergers prediction of two parties. Thus, Dickson and Scheve expect a non-linear relationship between group size and number of parties in restrictive systems a prediction obviously at odds with Duverger and those who followed. Using cross-national data on presidential election results and ethnic group demographics, they find support for their model.

The Scheve Dickson threshold, as we will call it in this paper, is defined by an underlying variable (A) that is a ratio of the size of the largest group relative to the size of the largest group plus the second largest group. The threshold is one when A is above 2/3, zero below 2/3.

11

As a final comment on social diversity and its impact on party systems, we note that several studies have observed that many ethnic groups have a nested structure: like Russian matroyshka dolls, large groups contain smaller groups, which break into still smaller groups, and so on down through the layers of a society (Scarritt and Mozaffar, 1999; Mozaffar, Scarritt, and Galaich, 2003; Posner, 2005; McLaughlin, 2007; Ferree, 2007). This is common but not restricted to Africa: Hispanics in the United States also break into many smaller groups Mexicans, Cubans, Salvadorans, etc. With a few notable exceptions, most studies of the effects of ethnic diversity on party systems fail to take nesting into account, measuring diversity at one level of the social structure and ignoring the others. This opens up the possibility of measurement error: a countrys ethno-linguistic fractionalization, or ELF, can be quite different, depending on whether it is calculated based on broad ethno-linguistic groupings or narrower tribal grouping. It also ignores theoretically interesting political dynamics created by nested structures. The payoff to exploring such dynamics is potentially large, as a handful of studies have shown. Mozaffar, Scarritt, and Galaich create a measure for overall ethnic diversity that aggregates diversity across levels of the nested structure. Brambor, Clark, and Golder (2006) use this measure to show that more diversity is associated with more parties. The intuition here seems to be a variant on the social diversity argument presented earlier: the more groups existing at all levels of the nested structure, the greater the demand for parties. In a recent paper on South Africa, McLaughlin (2007) offers an alternative argument. He suggests that different levels of a nested ethnic structure may actually exert opposing effects on the number of parties. In the South African case where African ethnic groups nest within a larger racial one, for example, he argues that racial diversity increases the number of parties in municipal elections while ethnic

12

diversity depresses it. According to this line of reasoning, where a large ethnic group nests within a dominant racial group there are two possible winning constituencies: the racial group and the ethnic group. A party (or parties) can arise to mobilize each of these constituencies. In contrast, when the dominant racial group is splintered into many small ethnic groups, none of which is a majority (or even close to one), the incentive for party entrepreneurs to mobilize one of these splinters is relatively low. As a result, we should see only one large racially based party. From this short review of the literature on social diversity, institutions and party systems, we obtain four additional hypotheses: Hypothesis 5: Social diversity increases the number of parties only when district magnitude is sufficiently large. In systems with low district magnitude (e.g. SMP systems), social diversity has no effect on the number of parties. See Clark and Golder 2006: 694. Hypothesis 6: Under plurality rule, the number of parties is greater above the Scheve- Dickson threshold versus below it. More technically, the number of parties is greater for 2/3 < A < 1 than for 1/2 < A < 2/3, where A is the size of the largest group divided by the size of the largest group plus the second largest group. See Dickson and Scheve 2007: 23. Hypothesis 7: In areas with nested groups, diversity on each level has complementary effects such that more total diversity (the sum of diversity at different levels of the nested structure) increases the number of parties. Hypothesis 8: In areas with nested groups, diversity on each level has countervailing effects, with some levels increasing the number of parties, and other levels decreasing them. 2. Prior Empirical Work and Its Limitations

13

As indicated above, numerous studies have tested hypotheses 1-8. However, many of these tests particularly those related to the interactive effects of social diversity and institutions possess flaws that make their findings less persuasive. First, many studies confront the problem of endogeneity. As Lipset and Rokkan (1967), Lijphart (1994), Ordeshook and Shvetsova (1994), and Boix (1999) have argued, the choice of institutions may not be independent of the nature of the existing party system. Countries with many political traditions may adopt PR rules as a means of accommodating diversity; countries with less political diversity may find plurality systems more palatable. Down the road, these choices may show up as a relationship between institutions and the party system, but we cannot say for sure which variable is causally prior. Many studies of party systems have not dealt with this problem in a satisfactory way. Second, many studies especially those looking at social diversity use cross-national datasets to test their sub-national theories. As Cox (1997) is at pains to show, Duvergers logic applies to districts, not countries. Duvergers Law could hold for all of the districts in a country but the country could still have more than two parties at the national level if the two parties winning in each district are different. This result, a failure of linkage (in Coxs words), can have many causes: the decentralization of political and economic power (Chhibber and Kollman 2004); federalism (Jones 1997, Samuels 2000); and the dispersion of power in national institutions (Hicken 2005). In spite of this wealth of knowledge, almost all studies of social diversity, institutions, and party system compare nations, not districts. While this may be due to a deficit of datasets that match social diversity with election results, it nonetheless undermines the credibility of these studies findings.

14

Third, cross-national studies of diversity and party systems suffer another problem, that of comparing apples and oranges. Most studies operationalize social diversity as ethnic diversity (e.g. Ordeshook and Shvetsova, 1994; Amorim Neto and Cox, 1997; Mozaffar, Scarritt, and Galaich, 2003; Clark and Golder, 2006; Brambor, Clark, and Golder, 2006; Dickson and Scheve, 2007). While ethnic diversity boasts the convenience of being collectable across countries, ethnicity is not relevant to politics in many nations. Indeed, there are many types of diversity -religious, economic, generational, regional, etc).that may be far more relevant to politics than ethnicity. Ignoring the different forms of social diversity across countries may not produce bias; however it is likely to introduce a great deal of noise in a study. Fourth, cross-national data limits the sample size tested. In many analyses, authors use only a single cross-section of countries, yielding well under 200 and sometimes less than 100 cases. Some studies attempt to circumvent this limitation by using repeated cross-sections (e.g. Mozaffar, Scarritt, and Galaich, 2003) but since ethnic diversity variables do not change over time, this strategy does not necessarily provide more information than single cross-sections. A few studies have avoided some or all of these problems by using sub-national data. Vatter (2003), using data on Swiss cantonal parliaments, finds evidence of an interaction effect between religious diversity and effective threshold. Lagos Penas (2004), looking regional electoral data from Spain, also finds evidence of an interaction effect between social diversity (here the strength of regional identities) and institutions. Geys (2006) utilizes a large dataset on Belgian municipal elections to demonstrate similar findings. We borrow this research strategy of using sub-national data in our analysis below. 3. Data and Tests

15

Our goal in this paper is to improve upon prior work by using better data to examine the relationship between institutions, diversity, and party systems. We do this by focusing on variation across local party systems in a single country, South Africa. As we discuss below, several unique properties of the South Africa data allow us to avoid many of the problems that have troubled earlier studies of party systems, allowing for greater confidence in our studys results. Since the end of apartheid, South Africans have voted in three sets of local (municipal) elections: 1995, 2000, and 2006. The first local elections were based on jurisdictional boundaries left over from apartheid.4 Subsequent to the 1995 elections, the ANC-led government embarked on an ambitious plan to re-draw local government boundaries to better reflect the new political dispensation (RSA 1998). It also allotted to these municipalities a completely new and important set of political and economic powers (RSA, 1998; RSA, 2004). It delegated to them the responsibility to provide almost all public services, with the exception of education and housing. It also attempted through direct elections for local councilors to create strong political accountability, and hence the incentive for municipal governments to provide these services. The result was a political map with 284 municipalities containing 3774 wards.5
4

Three basic systems of municipal government existed under the apartheid system, one for whites, another for blacks living in proximity to large white populations (i.e., blacks living in white South Africa), and a third for blacks living in black areas (called Bantustans or homelands). Coloureds and Indians also lived in racially segregated areas. These communities enjoyed more autonomy and better services than blacks but, generally speaking, had access to lower quality services than whites. White municipal governments provided local services to their municipality while the central government directly administered blacks who lived in white South Africa. Blacks living in homelands had nominal self-government because they lived in notionally independent states, but the apartheid regime supported Bantustan governments financially and intervened heavily in their affairs (de Beer and Lourens, 1995; Cloete, 1989; Wittenberg, 2003). The quality of local services differed across these three types of government. Whites received the highest level of services while blacks living in the impoverished homelands had almost no local services. The quality of services enjoyed by blacks living in white South Africa fell in between these two extremes. The most explosive protests against the system originated in the large black townships in white South Africa, often due to the glaring inequality in the living conditions of the black and white communities living in close proximity to one another.
5

The government divides municipalities into three categories, A, B, and C. Category A encompasses the six largest cities (Cape Town, Durban/eThekwini, East Rand/Ekurhuleni), Johannesburg, Port Elizabeth/Nelson Mandela, and

16

Given identical political borders in 2000 and 2006, we use the electoral data from both sets of local elections. We also exploit the data from the 2001 South African census, which used the same municipal boundaries. The institutional structure adopted for the 2000 and 2006 local elections offers a valuable opportunity for studying the effects of institutions on party systems. Following recent trends (e.g. Russia, Japan), South Africa adopted a mixed member electoral system: voters are represented on municipal councils by ward councilors as well as by at-large municipal councilors. The ward councilors are elected in ward elections under single member plurality (SMP) rules: several candidates compete for a single ward seat and the candidate winning the most votes gets the seat. In contrast, the at-large councilors are elected via proportional representation (PR) rules at the municipal level. Voters therefore cast both an SMP (ward) vote and a PR (municipal) vote. Both sets of election results are reported at the ward level, so we know how a particular ward voted in both the ward election and in the at-large PR election. By comparing these different voting outcomes, we can test the effects of institutions on the party system while holding constant the group of voters participating in the election. We are also able to match ward level election results with census data, which provides measures of racial and ethnic diversity and other socioeconomic indicators and allows us to investigate the duel effects of social cleavages and institutions on party systems. Thus, our dataset offers a number of important advantages. First, since the decision to adopt electoral institutions common to all wards occurred at the national level, and did not reflect social demographics or politics at the ward or municipal level, we avoid the problem of
Pretoria /Tshwane). Category B consists of all other inhabited areas. Category C municipalities have very small and widely scattered populations. The responsibilities of local governments vary by category. Category A municipalities have far more expansive responsibilities than those in Category B, and Category C municipalities possess the least powers (Gaffneys, 2004; IDASA, 2004; MDB, 2003; RSA, 2003).

17

endogeneity that plagued earlier work. We can therefore treat institutions as truly exogenous to our independent and dependent variables. Second, unlike studies that look at national party systems, we are looking at the equivalent of districts, which allows us to study Duvergers dynamics at the theoretically appropriate level of analysis. Third, unlike studies that pool together countries with significantly different levels of ethnic tension by looking at a sample from within a single country, we ensure that our units of analysis are more comparable. What varies across our sample is the size of different groups, not the types of groups being compared or the general significance of group differences. And fourth, by using wards as our units of analysis, we draw on a sample that contains over 3700 observations for each election, giving us much far more explanatory power than typical cross national studies. We should also highlight some possible caveats about our data. The first concerns the nature of political competition in South Africa. In all three national elections, the ANC dominated polling, winning more than 60 percent of the vote each time. The ANC also won elections in seven out of nine provinces. One might reasonably wonder if this dominance of the ANC so restricts variation in our dependent variable that we gain no traction in our tests. We do not believe this to be the case. Although the ANC clearly dominates at the national and provincial levels, it does not uniformly dominate at the ward level. In the 2000 election, the ANC won 69% of all wards (2577 out of 3715). The Democratic Alliance (DA) won 16%, the Inkhata Freedom Party (IFP) won 12%, and 16 other parties split up the remainder of the wards. In 2006, the ANC won 78% of all wards (3019 out of 3886), the DA 11%, the IFP 9%, and 11 other parties split up the rest. Hence, 22-31% of wards (at least 800 wards in each election) were won by a different party, giving us sufficient range of variation in our dependent variable to test our hypotheses.

18

A second caveat with regard to our data concerns the possibility of contamination effects between PR and ward races. Rather than assume that such mixed system that is, those that employ PR and SMD rules simultaneously -- provides researchers with a controlled experiment, for testing the effects on institutions on the number of parties, recent studies have found that parties run candidates in SMD elections to improve their chances in PR elections (Ferrara and Herron 2005, Ferrara, Herron and Nishikawa 2005, Herron and Nishikawa 2001). If parties tend to run candidates in non-competitive wards to pump up their municipal PR vote totals, we would expect that more parties competing in the municipal PR elections would result in higher effective number of parties in the ward races. Thus, to test for and hence control for the strategic choice of parties to enter SMP elections to affect the outcome of the PR one, we include the number of parties competing in the municipal PR election as an explanatory variable for the number of parties in SMP elections. In our analysis, we use two dependent variables, the effective number of parties (ENP) under SMP rules (voting for a ward representative) and ENP under PR rules (the ward level results of the municipality-wide PR election). We measure both dependent variables at the ward level for the 2000 and 2006 elections. Our independent variables consist of an upper bound (M+1), fractionalization measures, and relatively straightforward sociodemographic variables (see Table 1 for sample statistics). We review each in turn. [Table 1 Here] We generate our upper bound variable (Coxs M+1 measure) by adding one to the number of seats up for grabs in the PR election. As the PR elections occur at the municipal

19

election, and several wards comprise each municipality, the upper bound variable repeats across wards in the same municipality. This variable ranges from 4 to 109, with a mean of around 33, indicating fairly un-constraining bounds in most areas. Using the 2001 South African census, we created a measure of racial fractionalization (RF). South Africa has four major ethnic groups, Africans, coloureds, Indians, and whites. RF (akin to the commonly used ELF measure) is a Herfindahl index of the fraction of each racial group in the total population and indicates the probability that two individuals picked at random will be from different racial groups. 6 Higher values of RF indicate greater levels of fractionalization or diversity. The mean value for RF in our sample is .13, with a minimum value of 0 and a maximum value of .73. Figure 1 shows a kernel density plot of the racial fractionalization (RF) variable. As is evident from the graph, the majority of our wards are highly homogeneous. However, given the size of our dataset, we have a good spread of less homogeneous wards, especially up to an RF score of around .5. (Figure 1 here) Exploring the data to illuminate the RF measure, a score of zero or close to zero occurs when one racial group comprises all or nearly all of the population. The sixth ward of the Tswelopele (Hoopstad) municipality in Orange Free State exemplifies this: 98 percent of its population is African, 2 percent is coloured, which translates into an RF score of .04. This kind of African-majority ward is prevalent throughout most provinces of South Africa, with exceptions located in the Western Cape and Northern Cape, where coloureds are the dominant racial group and Africans are a minority. A score of around .5 usually indicates two well
6

The standard formula is: population.

K represents the number of groups and p is each groups fraction of the total

20

balanced racial groups. The Gauteng municipality of Midvaal (Meyerton), ward 9, provides an example of such a distribution. Here Africans are 44% of the population, whites are 55% of the population, and the RF score is .51. High RF scores correspond to situations where there are sizeable concentrations of at least three of the racial groups. These are relatively rare in our sample: we have only 84 wards where the RF is above .6. Our highest score occurs in the Durban (Kwa Zulu-Natal) area municipality of Newcastle, in the fourth ward. Here Africans comprise 30 percent of the population, Indians 33 percent, whites 23 percent, and coloureds 14 percent, translating to a RF score of .73. To test the possibility of interaction effects between racial fractionalization and institutions (specifically, the bounds), we also included an interaction of these two variables. In addition to the racial fractionalization score, we calculated the ELF measure using language groups (English, Afrikaans, isiNdebele, isiXhosa, isiZulu, Sepedi, Sesotho, Setswana, SiSwati, Tshivenda, Xitsonga, and other) as documented by the 2001 census. This variable ranges from 0 to .88, with a mean value of .31. Figure 2 shows a kernel density plot for the ELF variable. We have a large cluster of linguistically homogeneous wards (as with the RF variable) and a reasonable sampling of more heterogeneous wards. As suggested by the figures and mean values of RF and ELF, ELF tends to be higher than RF. The two variables have a correlation of . 44, reflecting the concentration of languages within racial groups. [Figure 2 here] Following McLaughlin (2007), for some of our specifications we include controls for the size of all of the groups comprising the RF and ELF indices. This ensures that any effects we attribute to fractionalization are a function of fractionalization, not of the size of the particular

21

groups in the fractionalization index. We also include variables for urbanization (measured at the level of municipality) and housing stock (percent living in formal housing, percent living in informal housing, and percent living in traditional housing) as these capture something of the socioeconomic flavor of the wards. Wards (urban and rural) with formal housing tend to be relatively well off and economically developed. Rural wards with traditional housing tend to indicate relatively less developed African wards in the countryside. Urban wards with informal housing tend to be poor city areas. We would expect information quality (important for coordination) to be highest in relatively more urbanized areas and areas with more formal housing. We also use dummy variables for each of the provinces. These dummy variables capture various unknown and unspecified factors that vary at the provincial level and might relate to both our dependent and independent variables and cause omitted variable bias if ignored. The provincial dummies also have substantively interesting interpretations. Successful challenges to the ANCs hegemony over the African vote have been geographically concentrated: in KwaZulu Natal, by the Inkatha Freedom Party (IFP); in the Eastern Cape, by the United Democratic Movement (UDM); and in the North West Province, by the United Christian Democratic Party (UCDP) For the most part, these African opposition parties which have grown up around politicians active in the Bantustan administrations of the apartheid era have been the only ones to draw even moderate portions of African voters away from the ANC. Dummy variables for these three provinces therefore capture (among other things) the presence of viable African challengers to the ANC and the possibility of coordination failure amongst African voters. We would expect (all else equal) the effective number of parties to be higher in these, especially in PR elections.

22

As mentioned, we need to control for the possibility of contamination effects whereby parties run candidates in SMP elections to boost their chances in the PR vote, which would lead to an inflated SMP ENP. To control for this effect, we include the number of parties competing in the municipal PR elections in our ward SMP specifications. Finally, we include the Scheve Dickson measure to test their threshold theory, i.e. whether low levels of fractionalization will encourage same-group competitors to enter the race.

4. Results To review, we expect ENP to be higher in the PR elections versus the SMP elections (Hypothesis 1). We expect ENP to correlate with the upper bound imposed by district size in PR elections (Hypothesis 2) and that ENP should not exceed the bounds imposed by the rules (district size plus 1 for PR elections, 2 for SMP elections) (Hypothesis 3). We expect racial diversity to have a much larger effect in PR elections versus SMP elections and we expect the size of diversity in PR elections to be conditioned by the upper bound imposed by district size (Hypothesis 5) and anticipate that ethnic diversity (controlling for racial diversity) could have either a positive or a negative effect on ENP (Hypotheses 7 and 8). We also test the threshold derived by Scheve and Dickson in their recent paper: they expect ENP be higher above this threshold (Hypothesis 6). And finally, anticipating that repeated elections might facilitate the coordination process behind Duvergers Law and Coxs M+1 Rule, we expect that the first three hypotheses should hold better in 2006 compared to 2000 (Hypothesis 4).

23

We begin by looking at simple comparisons of the effective number of parties (ENP) under SMP and PR rules for the 2000 and 2006 elections (Table 2). Contrary to the expectations of Hypothesis 1, we do not find significant differences in ENP under different institutional rules. For the 2000 elections, ENP hovered around 1.6 1.7 under both sets of institutional rules. The same was true in 2006. In 2000, the maximum effective number of parties was around 4.5 under both systems. In 2006, the maximum effective number of parties was higher for both systems, but especially PR, which reached around 5.5. [Table 2 Here] What is going on? One possible explanation lies in the low overall effective number of parties in these elections. With a mean well under two for both sets of rules, even restrictive bounds (such as occur in SMP systems) may not be exerting much downward pressure on party systems. In other words, if other factors are keeping the party system small (e.g. a dominant party like the ANC), institutional differences may not have the opportunity to emerge as important. If this is true, then the failure of Hypothesis 1 may not indicate the failure of Duvergers logic more broadly. Hypothesis 3, which rephrases Coxs M+1 rule, anticipates this possibility. If the primary explanation for the similarity in ENP under the different rules lies in factors which keep ENP below the bounds, then we should expect to find few instances (under either rules) of bound violations. Here again our results confound expectations. On the one hand, we see zero bound violations under PR systems: in every ward, for both elections, ENP was less than the number of seats up for grab in the district (M) plus one. On the other hand, we see massive violation of the M+ 1 rule under the SMP system: in 815 cases in 2000 (1015 in 2006), ENP exceeded 2. Thus,

24

around a quarter of the wards produced more parties than predicted during the ward (SMP) elections: a clear failure of Duvergers law. The failure of Hypothesis 1 is therefore not due solely to ENP being below the bounds for both systems. To explore the remaining hypotheses, we turn to a series of OLS models. We have strong reasons to believe that errors will correlate within municipalities: the wards within municipalities all participate in the same at-large PR election and are geographically close to one another. We therefore report robust standard errors clustered by municipality. Table 3 contains results for 2000, Table 4 for 2006. The first two columns of each table show results for ENP in the PR elections. The last three columns show results for ENP in the SMP elections. In the first four columns, we control for upper bound, racial fractionalization (RF) and ethnolinguistic fractionalization (ELF). In the second (PR) and fourth (SMP) columns we add in the full set of racial, ethnic, and provincial controls. The fifth column substitutes the Scheve Dickson threshold for the fractionalization measures. We have put in bold all results with p <.05. [Tables 3 and 4 here] Evaluating first the upper bound variable, we find no evidence suggesting that district size matters. The upper bound variable does not reach conventional significance levels for any of the 2000 or 2006 models. It is borderline significant for the simple model in 2006 (column 1), but the effect collapses when we include the full set of controls (group size, province, urban percentage, and housing stock).7 We suspect that the looseness of the bounds (mean is 33) reduces their impact on the party system. After all, the average ward would need to have demand for over thirty parties for the bounds to begin to have an effect!
7

We also experimented with using the natural log of the bounds. We found no evidence of any relationship between this modified variable and PR ENP.

25

In contrast, we do find evidence of contamination in our data. In both 2000 and 2006, the number of parties competing in the municipal election significantly correlates with the effective number of parties at the ward level. The size of the coefficient is modest but well estimated in most specifications: one extra party in the municipal election translates into an increase of around .02 effective number of ward parties. This suggests that South African parties may run extra candidates in ward elections in order to increase their vote totals in municipal elections. Turning to the demographic variables, we find strong and consistent evidence that racial fractionalization has a powerful impact on party systems. Moreover, this impact extends across institutional setups: for both PR and SMP elections, racial fractionalization greatly matters. For the 2000 elections, looking at the full models in columns 2 and 4, we see that a shift from the least fractionalized ward (RF = 0) to the most fractionalized ward (RF = 1) corresponds to a gain in 1.3 effective parties for PR rules and a gain in 1.2 effective parties for SMP rules. For the full models in 2006, the same shift results in a gain of .6 effective parties for PR rules and .7 effective parties for SMP rules. Moreover, we note that the interaction variable between municipal bounds and racial diversity never approached significance in any of our models. Thus, contrary to Hypothesis 5, the effects of racial fractionalization do not depend on institutional framework. The distribution of racial groups appears to drive outcomes in both PR and SMP systems. To evaluate whether non-linearities or thresholds might condition the relationship between racial fractionalization and the effective number of parties, we generated a simple scatterplot of the two variables for 2000 (Figure 3, based on data from SMP rules). If Duvergers logic applies, we might expect to see a correlation between RF and ENP where ENP

26

is less than 2 (the bounds imposed by the SMP rules), and then a truncating of ENP after that level regardless of RF. However, as the plot makes clear, there is a basic linear relationship between RF and ELF and no evidence of threshold points, non-linear areas, or truncation: At low levels of fractionalization, we tend to see only one party (the dark clump of cases in the lower left corner of the plot show that this quite common). As fractionalization nears .5, we see two parties. And as fractionalization moves beyond .5, we see increasing cases of more than two parties. Of course, there are many outliers to the relationship. Even at very low levels of RF we see violations of the bounds, suggesting other factors matter as well. [Figure 3 Here] The size of different racial groups is clearly one of these factors, as evidenced by the coefficients on percent African, percent coloured, and percent Indian (percent white is the reference category). Holding constant racial fractionalization, wards with large concentrations of Indians and/or coloureds have significantly more parties than white and/or African wards. The coefficients are large, well estimated and consistent across institutions and years. In 2000, a district that was 100 percent coloured had about 1 more effective party than one that had no coloureds (under both rules). In 2006, the effect was .6-.7. The presence of Indians had an even larger effect, around 1.7 in 2000, 1.2-1.3 in 2006. We suspect there are two factors at work. First, per earlier research, coloured and Indian voters tend to support a wider range of parties than Africans and whites (Schulz-Herzenberg 2007). Furthermore, these groups behavior is less predictable, which makes it harder for voters to form common beliefs about the relative strengths of parties, which in turn increases the difficulty of Duvergian coordination and leads to more parties. In contrast, the presence of Africans tends to decrease the number of parties relative to all other groups: in 2000 the coefficient was around -.7 to -.8 for both systems, and in 2006 it

27

was about -1.3. We suspect that this occurs because, outside of a few provinces, African support for the ANC is very solid. This has a direct dampening effect on the number of parties and an indirect effect as well because it reduces uncertainty for other groups and makes coordination easier. Turning to the remaining variables, we find suggestive evidence that ethnolinguistic fractionalization affects ENP. In all specifications, the coefficient on ELF was negative confirming earlier work by McLaughlin (2007). In 2000, in the basic model (columns 1 and 3) it was statistically significant at the .05 level for both institutional systems, although the size of the coefficient was rather modest, and not significant in the full model. A similar pattern prevailed in 2006, except the coefficients were smaller. Interestingly, if we leave out racial fractionalization, then ELF has a positive effect. Altogether, the ELF results suggest that holding constant racial demographics ethnic homogeneity promotes party proliferation, but only in relatively minor and not particularly robust ways. Racial fractionalization has a much larger and systematic impact on parties in South Africa. This is not surprising given the low level of salience ethnic divisions have generally had in recent South African politics compared to racial ones. The size of different language groups also exerts an independent effect on ENP. The reference category are isiZulu speakers, and the results show consistently that areas with more isiXhosa, isiZulu and Setswana speakers have somewhat higher ENP relative to areas with Sepedi, SiSwati, Sesetho, Tsonga, Ndebele, English, and Afrikaans speakers, although most of the effects are rather small.8 The effect of concentrations of Sesotho and Tshivenda speakers is
8

Recall that these effects are holding race constant, hence the coefficients on Afrikaans and English speakers are probably picking up differences between Africans. They suggest that areas with higher percentages of Africans who speak either Afrikaans or English (not a large group) are associated with fewer parties.

28

inconsistent, leading to (more parties in 2000 and less parties in 2006). It is interesting to note that the Xhosa, Zulu and Tswana all had separate homelands under apartheid, as did the Venda. Most of the other groups (the Sotho are the exception) lived in more heterogeneous patches of the northern and eastern Transvaal. The greater number of parties associated with the Xhosa, Zulu and Tswana may reflect as earlier suggested that small African opposition parties have grown around former homeland politicians. Once the size of the various ethnic and racial groups is taken into account (as they are in Tables 3 and 4), provincial dummies have less significance and have inconsistent effects across elections. The Eastern and Northern Cape had fewer parties both years. In 2000, the Western Cape and Mpumalanga also had fewer parties. In 2006, KwaZulu-Natal had more. If the variables for the size of racial and ethnic groups are removed -- leaving in the fractionalization indices -- then Western Cape and KwaZulu-Natal both have very large, very well estimated positive coefficients. The North West Province and the Northern Cape also have positive coefficients, although they are not as large. The provincial dummies in this case are picking up the effects of coloured and Indian voters (concentrated in the Western and Northern Capes and KwaZulu-Natal) as well as the presence of Zulu and Tswana groups (in KwaZulu-Natal and the North West province). A few final observations: our results are inconsistent with regard to the environment variables percent urban, percent living in formal housing, and percent living in traditional housing. In 2000, more urban areas apparently had fewer parties. Housing stock did not matter. In 2006, percent urban had no effect, but areas with more formal housing had more parties. We do not have a substantive story for these effects.

29

As for the Scheve-Dickson hypothesis, their threshold variable is highly significant for both elections, but it has the wrong sign: ENP is substantially lower above the threshold. When we recall that areas above the threshold are ones where the largest group is relatively dominant (specifically, when the ratio of the largest group to the largest group plus the second largest group is above two thirds), then it becomes clear that the Scheve-Dickson threshold simply confirms the earlier result: that more homogeneous areas have fewer parties than less homogeneous areas. This does not imply that same group challengers never enter races. Several parties have, for example, vied for the African vote. However, what seems to determine the entrance of these parties into the race is not the size of the dominant group but institutional legacy. All of these parties (the IFP, UDM, UCDP) have arisen around old homeland parties and personalities. Finally, we find no evidence that Duvergers logic holds better in the second election. Institutional effects were as weak in 2006 as they were in 2000. Our models fit somewhat better in 2006 (as measured by R2), but we hesitate to place much emphasis on this. On the whole, we find little support for Hypothesis 4. 5. Summary and Discussion To summarize our results: we find little evidence for institutional effects. Our data show no difference in the effective number of parties for SMP versus PR rules. We do find a moderate relationship between municipality bounds and the effective number of parties, as expected, but this relationship is not especially robust and disappears when we control for percent urban and include provincial and racial dummies perhaps because the bounds are so large in most areas that they do not have much of a constraining effect on parties. More

30

significantly, we observe widespread failure of Duvergers Law in the SMP elections: about a quarter of our wards had more than two effective parties. We find no evidence that social diversity matters more under PR rules since the effects of racial and ethnic fractionalization hold across electoral institutions. Finally, our results pertain both to the earlier (2000) and later (2006) elections, providing no evidence of institutionalization over this period In contrast, we find strong evidence that social cleavages drive South Africas local party systems. Racial fractionalization has a large, consistent, and statistically significant effect on the effective number of parties. This holds across elections, across institutional setups, and across configurations of control variables. Furthermore, the relationship between racial fractionalization and ENP in the local South African data is linear: low levels of fractionalization translate into party hegemony (ENP scores well under two), high levels of fractionalization push the party system beyond the Duvergian bounds. This finding contradicts Scheve and Dicksons intuition that low levels of fractionalization will encourage same-group competitors to enter the race. In addition to racial fractionalization, we also find that the presence of some groups (coloureds and Indians) increases party number, while the presence of others (Africans) depresses it. Confirming the recent work of McLaughlin (2007) we find that ethnic fractionalization shapes South African local party systems. The effect is much smaller than that of racial fractionalization and is sensitive to specification, but is consistent across models, institutions, and years: holding constant RF, ELF appears to decrease the number of parties. We also find that some ethnic groups are consistently associated with higher numbers of parties (Zulu, Xhosa, Tswana), perhaps reflecting the tendency of African challengers to the ANC to grow around old

31

homeland administrations and personalities. After controlling for ethnic groups, provincial dummy variables have relatively minor and inconsistent effects. Some of our results warrant greater discussion. In particular, how do we explain the linear relationship between racial fractionalization and the effective number of parties under SMP rules? These results confound Duvergian logic and contradict the prior work of Ordeshook and Svetlova, Amorim-Neto and Cox, and others who anticipate a truncation of the number of parties at around two in SMP systems, regardless of social cleavage structure. One explanation is that South Africas voters and politicians are primarily motivated by expressive concerns (see Horowitz 1985 for a discussion of expressive voters). Cox (1997) shows that successful coordination depends on a number of conditions, one of which is that voters are short-term, instrumentally rational. When voters do not conform to this assumption when their primary desire is to express their identity and allegiance with a party they may decide to stick with their first choice political party (or candidate) rather than moving to their second (or third or beyond) choice when it appears obvious that their first choice cannot win. In other words, expressive voters are likely to be sincere voters, and sincere voters do not conform to Duvergers logic. This could account for the bound failures and the unexpected correlation between racial fractionalization and ENP under SMP rules in our data. However, this explanation contradicts Mattes (1995) and Ferree (2007), who find little evidence of expressive voting in South Africa in survey data.9
9

Building on Mattes (1995), Ferree (2007) argues that South African voters use the racial credentials of parties as an informational shortcut to predict who they will favor when in office. Africans have continued to support the ANC not out of a desire to express identity, but because they view the existing opposition as white and therefore discount its credibility. Until the current opposition transforms its image (a difficult process at best) or alternative opposition parties with a national presence arise, the ANC is likely to continue to monopolize the African vote. If this argument is correct, it suggests that African behavior is not driven by expressive desires and hence should be responsive to strategic considerations. Of course, African voting alone does not determine our results. Even if African voters uniformly support the ANC, Duvergers prediction would still hold at the ward level if the other three groups coordinated on a single party.

32

A second explanation focuses on the information requirements of Duvergers Law and how social diversity might impact them. Cox (1997) shows that, in order to coordinate, voters must have a clear, commonly understood ranking of parties. Absent this, voters cannot determine their first choices probability of winning and have no reason to alter their behavior for strategic purposes. High levels of social fractionalization produce precisely this sort of situation: High fractionalization generally indicates the presence of multiple similarly sized groups. Assuming each group has a favorite party, this means there are multiple similarly sized parties and no clear ordering of winners and losers. Voters, quite rationally, have no reason not to support their first choice. South Africas unique political demographics produce a further twist on this situation. Two of the countrys four racial groups coloureds and Indians are known to support a wide range of parties. Indians have an Indian party to support (the Minority Front), but only a minority of Indians actually vote for this party. Hence, when these voters are present, it is especially difficult for voters to come up with a rank ordering of the parties. And high fractionalization scores necessarily involve a large concentration of one or more of these groups. An example from the data helps demonstrate these dynamics. Earlier we discussed the fourth ward of the Newcastle municipality (near Durban, in KwaZulu Natal). This ward has our highest RF score (.73). It is 30 percent African, 14 percent coloured, 33 percent Indian, and 23 percent white. Even if we assume whites and Africans each have their own party, voters will still be unable to rank order the parties because of uncertainty about the likely behavior of coloureds and Indians. In this situation, voters have no reason to behave strategically and would stick with their first choice not because they are not short-term instrumentally rational, but because they lack the information to behave differently. This reasoning is born out in this ward:

33

ENP was around 2.7 in both elections. In 2000, the Democratic Alliance won with 51 percent of the vote, the African National Congress took 30 percent, the Inkatha Freedom Party 14 percent, and the Minority Front 5 percent. In 2006, the DA won 54 percent of the vote, the ANC 28 percent, the IFP 4 percent, and five additional parties split up the remainder. Which of these accounts expressive voting or the lack of information best accounts for the failure of Duvergers Law in South African local elections? We suspect both factors matter. In either case, the South African data suggest that the institutions/social diversity interaction story prevalent in the literature is too simple. Where we see social diversity, we also see conditions that are likely to violate the assumptions necessary for Duvergers Law to hold. Hence, we are likely to see social diversity affecting party systems even under restrictive electoral institutions. 6. Conclusion The debate between the role of institutional and social factors on political outcomes will continue to remain a lively source of debate. In this paper we advance this research agenda by utilizing a sub-national dataset from South Africa that allows us to avoid several of the problems encountered by earlier studies looking at the relationship between electoral institutions, social diversity, and party systems. Specifically, because the choice of electoral systems at the local level in South Africa and the racial/ethnic mix of their populations were exogenous to local party systems, we avoid the problem of endogeneity that plagued earlier studies. Furthermore, by focusing on wards, we are able to test Duvergers intuitions at the theoretically appropriate level of analysis.

34

Our results confirm some conventional findings and cast doubt on others. Like earlier studies, we find that some forms of diversity (in this case, racial) increase the number of parties in the party system. However, we also find that other forms of diversity (in this case, ethnic) decrease the number of parties. Our results suggest that previous work, which has focused on just one type of diversity, might miss these theoretically interesting dynamics. Furthermore, unlike earlier studies, we find no evidence that electoral institutions mediate the effects of diversity. At least for South African local elections, diversity matters even in majoritarian elections. The explanation for this might lie in the effects ethnic diversity has on strategic behavior: it might either reduce such behavior (if voters are expressive, for example) or make such behavior difficult by obscuring common knowledge about the ranking of parties. If ethnic diversity can matter in majoritarian elections, why does this effect show up in South African elections but not elsewhere in cross-national studies, for example, or in Geys (2006) study of Belgian local elections? While it is impossible to say for sure, we suspect that the youth of South Africas democracy might play a role, especially if the problem lies in voters ability to rank the likely finishing order of parties. Whereas Belgian voters have had close to a century to figure out how to coordinate, South African voters have had a little more than a decade. In keeping with this conjecture, Clark and Golder (2006) note that the interactive effect between diversity and institutions is strongest in their data for the long-term democracies. Future work that considers sub-national dynamics in new democracies may shed more light on these relationships.

35

Tables and Figures

Table 1: Sample Statistics (Independent Variables)

Observations Upper bound Percent African Percent Coloured Percent Indian Percent White RF ELF Percent Formal Housing Percent Traditional Housing Percent Informal Housing Percent Urban Scheve Dickson Threshold 3760 3714 3714 3714 3714 3714 3714 3714 3714 3714 3760 3714 Mean 33 .80 .09 .01 .09 .13 .31 .67 .20 .13 .49 .92 Standard deviation 28 .33 .23 .08 .20 .19 .26 .27 .27 .18 .36 .28 Minimum 4 0 0 0 0 0 0 .01 0 0 0 0 Maximum 109 1 .99 .98 .96 .73 .88 1 .96 .99 .99 1

36

Table 2: Basic comparisons (ENP) 2000 SMP 1.66 4.53 815 3690 PR 1.63 4.53 0 3759 2006 SMP 1.70 4.86 1015 3884 PR 1.65 5.49 0 3895

Mean Maximum M+1 Violations Observations

37 Table 3: 2000 Results

(1) PR ENP 0.000 (0.001) -0.002 (0.002) (2) PR ENP 0.001 (0.001) -0.002 (0.002) (3) SMP ENP (4) SMP ENP (5) SMP ENP

Bound Bound/Racial Diversity Muni. Number of Parties Racial Diversity Ethnic Diversity Scheve Dickson Threshold Percent Urban Percent Trad. Housing Percent Formal Housing Percent African Percent Coloured Percent Indian Percent English Percent Afrikaans Percent isiNdebele Percent isiXhosa Percent other language Percent Sepedi Percent Sesotho Percent Setswana Percent Siswati Percent Tshivenda Percent Xitsonga WC

1.388** (0.101) -0.429** (0.085)

1.253** (0.128) -0.121 (0.063)

0.006 (0.008) 1.228** (0.091) -0.401** (0.090)

0.023** (0.009) 1.161** (0.103) -0.091 (0.065)

0.002 (0.009)

-0.222** -0.230** (0.084) 0.030 (0.080) 0.105 (0.071) -0.842* (0.375) 1.036** (0.093) 1.697** (0.137) -1.543** (0.446) -1.223** (0.390) -0.294** (0.106) 0.027 (0.131) 0.459 (0.741) -0.391** (0.077) -0.097 (0.142) 0.018 (0.110) -0.443** (0.071) -0.252 (0.133) -0.201* (0.092) -0.292** (0.098) -0.292** (0.087) -0.020 (0.075) 0.092 (0.068) -0.679* (0.270) 0.999** (0.107) 1.719** (0.145) -1.555** (0.350) -0.960** (0.293) -0.147 (0.112) 0.055 (0.132) 0.391 (0.665) -0.407** (0.080) 0.043 (0.122) 0.083 (0.108) -0.458** (0.092) -0.411** (0.131) -0.193* (0.089) -0.299** (0.102) (0.033) -0.151 (0.088) -0.066 (0.082) 0.002 (0.070) -0.126 (0.071) 0.784** (0.105) 1.312** (0.202)

-0.039 (0.097)

38
EC -0.265* (0.109) NC -0.386** (0.100) FS -0.080 (0.101) KZN -0.024 (0.085) NW -0.157 (0.090) GT -0.068 (0.076) MPUMA -0.140** (0.045) Constant 1.578** 2.545** (0.037) (0.392) Observations 3787 3787 R-squared 0.22 0.37 Robust standard errors in parentheses, clustered by municipality. * significant at 5%; ** significant at 1% -0.225* (0.112) -0.434** (0.100) -0.166 (0.088) 0.005 (0.084) -0.158 (0.093) -0.155* (0.067) -0.126** (0.044) 2.326** (0.285) 3761 0.33 0.058 (0.091) -0.165 (0.099) 0.117 (0.077) 0.292** (0.072) 0.172* (0.067) 0.061 (0.075) -0.066 (0.062) 1.869** (0.132) 3416 0.27

1.579** (0.039) 3761 0.18

39

Table 4: 2006 Results

(1) PR ENP 0.002 (0.001) -0.002 (0.004) (2) PR ENP 0.000 (0.001) 0.001 (0.004) (3) SMP ENP (4) SMP ENP (5) SMP ENP

Bound Bound/Racial Diversity Muni. Number of Parties Racial Diversity Ethnic Diversity Scheve Dickson Threshold Percent Urban Percent Trad. Housing Percent Formal Housing Percent African Percent Coloured Percent Indian Percent English Percent Afrikaans Percent isiNdebele Percent isiXhosa Percent other language Percent Sepedi Percent Sesotho Percent Setswana Percent Siswati Percent Tshivenda Percent Xitsonga

1.680** (0.132) -0.578** (0.074)

0.568** (0.144) -0.058 (0.096)

0.020** (0.005) 1.524** (0.107) -0.618** (0.084)

0.015** (0.006) 0.704** (0.171) -0.096 (0.104)

0.006 (0.005)

-0.194** -0.026 (0.087) 0.078 (0.094) 0.176* (0.076) -1.171** (0.293) 0.591** (0.210) 1.262** (0.272) -1.170** (0.416) -0.968** (0.289) -0.127 (0.111) 0.020 (0.144) 0.580* (0.281) -0.247* (0.103) -0.224 (0.114) -0.008 (0.113) -0.417** (0.081) -0.258* (0.113) -0.121 (0.131) -0.064 (0.083) 0.104 (0.083) 0.229** (0.077) -1.262** (0.324) 0.686** (0.201) 1.165** (0.259) -1.497** (0.415) -1.054** (0.327) 0.112 (0.133) 0.099 (0.138) 0.523 (0.271) -0.233* (0.110) -0.119 (0.112) 0.202 (0.123) -0.438** (0.084) -0.259* (0.113) -0.112 (0.130) (0.042) 0.030 (0.086) 0.117 (0.085) 0.191* (0.080) -0.368** (0.083) 0.578** (0.166) 0.919** (0.257)

40
WC 0.161 (0.136) EC -0.238* (0.108) NC -0.307* (0.139) FS 0.062 (0.079) KZN 0.290** (0.106) NW -0.086 (0.067) GT 0.011 (0.084) MPUMA 0.006 (0.077) Constant 1.548** 2.586** (0.048) (0.318) Observations 3409 3409 R-squared 0.24 0.45 Robust standard errors in parentheses, clustered by municipality. * significant at 5%; ** significant at 1% 0.052 (0.144) -0.263* (0.120) -0.396** (0.145) 0.016 (0.091) 0.311** (0.110) -0.148 (0.093) -0.108 (0.093) 0.014 (0.088) 2.567** (0.348) 3406 0.40 0.246* (0.104) -0.048 (0.057) -0.226* (0.114) 0.075 (0.068) 0.453** (0.073) 0.146* (0.068) -0.006 (0.073) 0.018 (0.064) 1.773** (0.157) 3108 0.38

1.529** (0.044) 3406 0.21

41

Figure 1: Kernel Density Plot of Racial Fractionalization.

0 0

Density 4

.2

.4 ELF_race

.6

.8

42

Figure 2: Kernel Density Plot of Ethnolinguistic Fractionalization

0 0

.5

Density 1.5

2.5

.2

.4 .6 ELF_language

.8

43

Figure 3: Scatterplot of ward (SMP) ENP and racial fractionalization (RF) for 2000.

1 0

ward_enp 3

.2

.4 ELF_race

.6

.8

44

References Amorim-Neto, Octavio, and Gary W. Cox. 1997. Electoral Institutions, Cleavage Structures, and the Number of Parties. American Journal of Political Science 41(1): 149-174. Benoit, Kenneth R. 2001. District Magnitude, Electoral Formula, and the Number of Parties. European Journal of Political Research 39(2): 203-224. Synergy, ISI Benoit, Kenneth. 2006. Duvergers Law and the Study of Electoral Systems. French Politics: 69-83. Boix, Carles. 1999. Setting the Rules of the Game: The Choice of Electoral Systems in New Democracies. American Political Science Review 93(3): 609-624. Brambor, Thomas, William Clark, and Matt Golder. 2006. Are African Party Systems Different? Electoral Studies 26: 315-323 Buhlungu, Sakhela, John Daniel, Roger Southall, Jessica Lutchman: HSRC Press, Cape Town, South Africa Chandra, K. 2004. Why Ethnic Parties Succeed: Patronage and Ethnic Headcounts in India. Cambridge: Cambridge University Press. Chhibber, Pradeep and Ken Kollman. 2004. The Formation of National Party Systems: Federalism and Party Competition in Canada, Great Britain, India, and the United States. Princeton: Princeton University Press.

45

Clark, William Roberts and Matt Golder. 2006. Rehabilitating Duvergers Theory: Testing the Mechanical and Strategic Modifying Effects of Electoral Laws. Comparative Political Studies 39(6), August: 679-708. Cloete, J.N. 1989. South African Local Government Administration. Pretoria, South Africa: JL van Schaik. Cox, Gary. 1997. Making Votes Count. Cambridge: Cambridge University Press. de Beer and Lauren. 1995. Dickson, Eric, and Kenneth Scheve. 2007. Social Identity, Electoral Institutions, and the Number of Candidates. Yale University: Typescript. Duverger, Maurice . 1963. Political Parties. New York: Wiley. Ferrara, F., and Erik Herron. 2005. Going It Alone? Strategic Entry under Mixed Electoral Rules. American Journal of Political Science 49(1): 16-31 Ferrara, F., E.S. Herron, and M. Nishikawa. 2005. Mixed Electoral Systems: Contamination and Its Consequences. New York: Palgrave Macmillan Press (St. Martins Press) Ferree, Karen. 2006. Explaining South Africas Racial Census. Journal of Politics, 68(4), November: 803-815. Ferree, Karen. 2007. The Political Origins of South Africas Racial Census. Unpublished book manuscript. Gaffneys. 2004. Gaffneys Local Government in South Africa. Johannesburg: Gaffneys Group.

46

Geys, Benny. 2006. District Magnitude, Social Heterogeneity and Local Party System Fragmentation. Party Politics 12: 281 Herron, Erik and M. Nishikawa. 2001. Contamination Effects and the Number of Parties in Mixed Superposition Electoral Systems. Electoral Studies 20: 63-86 Herron, Erik and M. Nishikawa. Mixed Electoral Rules' Impact on Party Systems. Electoral Studies 23: 753-768 Hicken, Allen. 2005. Engineering Aggregation? Institutional Reform and Party Aggregation in Thailand. University of Michigan: unpublished manuscript. Horowitz, Donald. 1985. Ethnic Groups in Conflict. Berkeley, CA: University of California Press, pg. 57 IDASA. 2004. Local Government Powers and Functions (Revised Version). Pretoria: Institute for Democracy in South Africa. Jones, Mark. 1997. Federalism and the Number of Parties in Argentine Congressional Elections. Journal of Politics 59(2): 538-549. Laakso, M., and Taagepera, R. 1979. Effective Number of Parties: A Measure with Application to West Europe. Comparative Political Studies 12: 3-27. Lago Penas, I. 2004. Cleavages and Thresholds: the Political Consequences of Electoral Law in the Spanish Autonomous Communities, 1980-2000. Electoral Studies 23: 23-43. Lijphart, A. 1994. Electoral Systems and Party Systems. Oxford University Press.

47

Lipset, Seymour and Stein Rokkan. 1967. Cleavage Structures, Party Systems, and Voter Alignments: An Introduction, in Lipset and Rokkan (eds.), Party Systems and Voter Alignments: Cross-National Perspectives. New York: Free Press. Mayes. 1995. McLaughlin, Eric S. 2007. Beyond the Racial Census: the Political Salience of Ethnolinguistic Cleavages in South Africa. Comparative Political Studies 40 Moser, Robert G. 1999. Electoral Systems and the Number of Parties in Post-communist States. World Politics 51: 359-384. Mozaffar, Shaheen, James Scarritt, and Glen Galaich. 2003. Electoral Institutions, Ethnopolitical Cleavages, and Party Systems in Africas Emerging Democracies. American Political Science Review 97(3): 379-390. Municipal Demarcation Board, Republic of South Africa. 2003. Local Government Functions: Definitions of Norms and Standards. Pretoria: Municipal Demarcation Board, Government of the Republic of South Africa. Ordeshook, Peter and Olga Shvetsova. 1994. Ethnic Heterogeneity, District Magnitude, and the Number of Parties. American Journal of Political Science 38(1): 100-123. Posner, Daniel N. 2005. Institutions and Ethnic Politics in Africa. Cambridge University Press (Political Economy of Institutions and Decisions series). Powell, G. B. 1982. Contemporary Democracies: Participation, Stability and Violence. Cambridge, MA: Harvard University Press.

48

Riker, William H. 1976. "The Number of Political Parties: A Reexamination of Duverger's Law." Comparative Politics 9(1): 93-106 Riker, William H. 1982. "The Two-Party System and Duverger's Law: An Essay in the History of Political Science." American Political Science Review 76(4): 753-766 RSA. 1998. The White Paper on Local Government. Pretoria: Ministry for Local Government and Provincial Affairs, Government of the Republic of South Africa. _____. 2003. Local Government Fact Book. Pretoria: Department of Provincial and Local Government, Government of the Republic of South Africa. _____. 2004. Local Government Municipal Rates Act 2004. Cape Town: Parliament of the Republic of South Africa. Samuels, David. 2000. The Gubernatorial Coattails Effect: Federalism and Congressional Elections in Brazil. Journal of Politics 62(10): 240-253. Scarritt, James R. and Shaheen Mozaffar. 1999. The Specification of Ethnic Cleavages and Ethnopolitical Groups for the Analysis of Democratic Competition in Africa. Nationalism and Ethnic Politics 5 (Spring): 82117. Schulz-Herzenberg, Collette. 2007. A Silent Revolution: South African voters, 1994 2006. State of the Nation: South Africa Taagepera, R. and Shugart, M. 1989. Seats and Votes: The Effects and Determinants of Electoral Systems. Connecticut: Yale University Press. Taagapera and Shugart. 1993.

49

Taagapera. 1999. Vatter, Adrian. 2003. Legislative Party Fragmentation in Swiss Cantons: A Function of Cleavage Structures or Electoral Institutions? Party Politics 9(4): 445-461 Wittenberg. 2003. Decentralization in South Africa. Working Paper. University of Witwatersrand, Johannesburg, South Africa.